Adaptive catalyst design based on high-throughput experimentation and machine learning

One of the advantages of solid catalysts is their multi-component design, which allows for catalytic multifunctionality. However, the way in which the components interact with each other and influence the catalysis is complicated. Therefore, such multi-component design is difficult to generalize, and have relied heavily on trials and errors or the sense of the individual researchers. A data-driven approach using machine learning has a potential to transform this sense-driven approach, provided that sufficient quantity and quality of data is available. Here, we present adaptive catalyst design that loops through high-throughput experimentation, machine learning-based catalyst recommendation, and active learning.
The catalysts comprising the training data were randomly selected from a pre-defined materials space. They were prepared by parallelized wet impregnation and evaluated in a home-made high-throughput screening system. The system automatically acquires the performance of 20 catalysts under a programmed series of reaction conditions, leading to a maximum of about 4,000 data points per day. Machine was trained on the acquired catalyst data, and it was used to recommend catalysts that were assumed to have high performance, or whose performance could not be assumed from the training data. These catalysts were prepared and evaluated to augment the training data. By looping through the above, the machine learning model evolves to accommodate a wider range of combinations, and eventually leads to the discovery of high-performance catalysts, in a sense-free manner.
The concept will be demonstrated on different types of catalysis for delivering a forefront of catalyst informatics.